算力网络架构与场景分析

Analysis of Computation Network Architecture and According Scenarios

为了发挥分布式部署的算力资源池化优势,需要构建一张用于连接这些算力资源的网络,打破孤岛壁垒,为用户应用提供最优算力服务,并最优化全网算力资源利用率。算力网络架构设计必须在应用的算力需求满足和网络改造部署成本之间保持均衡。文章重点分析了集中式算力网络架构、分布式算力网络架构和混合式算力网络架构三种方案在收敛速度、全网资源视图、应用算力资源映射、设备和协议升级等方面的不同特征,并描述各自适用的应用场景。经分析,混合式架构结合了分布式与集中式架构的优势,在实际网络部署中有更大的参考意义。此外,在新型的网络路由架构(比如ICN)下,算力网络将获得更加优异的调度和转发性能。

To take advantage of pooling benefits of the distributed computing resources in the network,it’s necessary to specifically design a network to connect these resources with the isolated resource barriers broken and service quality optimized upon the basis of the whole computation network.It’s extremely important to strike a good balance between satisfying service computation requirements and cost of upgrading the devices and protocols.This paper focuses on the analysis upon three architecture solutions in terms of convergence speed,whole picture of network computation resource,mapping between service and computation resources and upgrading of devices and protocols,as well as the scenarios best suitable under each solution.Distributed architecture boasts of faster convergence speed while comes with higher cost for device and protocol upgrading,centralized architecture holds the benefit of whole view picture of the network computation resource and limited impacts upon the devices and protocols.The hybrid architecture combines the advantages of the two solutions and tries to minimize the disadvantages.In addition,under the new network routing architecture,the computational power network would obtain more excellent scheduling and forwarding performance.